def test_filter(self):
raw_data = create_random_data(100)
tfexamples = list(map(preprocessing.make_tf_example, *zip(*raw_data)))
with tempfile.NamedTemporaryFile() as f:
preprocessing.write_tf_examples(f.name, tfexamples)
recovered_data = self.extract_data(f.name, filter_amount=.05)
self.assertLess(len(recovered_data), 50)
def gather(selfplay_dir, training_chunk_dir, params):
"""Gather selfplay data into large training chunk.
Args:
selfplay_dir: Where to look for games. Set as 'base_dir/data/selfplay/'.
training_chunk_dir: where to put collected games. Set as
'base_dir/data/training_chunks/'.
params: A MiniGoParams instance of hyperparameters for the model.
"""
# Check the selfplay data from the most recent 50 models.
_ensure_dir_exists(training_chunk_dir)
sorted_model_dirs = sorted(tf.gfile.ListDirectory(selfplay_dir))
models = [
model_dir.strip('/')
for model_dir in sorted_model_dirs[-params.gather_generation:]
]
with utils.logged_timer('Finding existing tfrecords...'):
model_gamedata = {
model: tf.gfile.Glob(
os.path.join(selfplay_dir, model, '*' + _TF_RECORD_SUFFIX))
for model in models
}
print('Found {} models'.format(len(models)))
for model_name, record_files in sorted(model_gamedata.items()):
print(' {}: {} files'.format(model_name, len(record_files)))
meta_file = os.path.join(training_chunk_dir, 'meta.txt')
try:
with tf.gfile.GFile(meta_file, 'r') as f:
already_processed = set(f.read().split())
except tf.errors.NotFoundError:
already_processed = set()
num_already_processed = len(already_processed)
for model_name, record_files in sorted(model_gamedata.items()):
if set(record_files) <= already_processed:
continue
print('Gathering files from {}:'.format(model_name))
tf_examples = preprocessing.shuffle_tf_examples(
params.shuffle_buffer_size, params.examples_per_chunk,
record_files)
# tqdm to make the loops show a smart progress meter
for i, example_batch in enumerate(tf_examples):
output_record = os.path.join(training_chunk_dir,
('{}-{}' + _TF_RECORD_SUFFIX).format(
model_name, str(i)))
preprocessing.write_tf_examples(output_record,
example_batch,
serialize=False)
already_processed.update(record_files)
print('Processed {} new files'.format(
len(already_processed) - num_already_processed))
with tf.gfile.GFile(meta_file, 'w') as f:
f.write('\n'.join(sorted(already_processed)))
def test_serialize_round_trip(self):
np.random.seed(1)
raw_data = create_random_data(10)
tfexamples = list(map(preprocessing.make_tf_example, *zip(*raw_data)))
with tempfile.NamedTemporaryFile() as f:
preprocessing.write_tf_examples(f.name, tfexamples)
recovered_data = self.extract_data(f.name)
self.assertEqualData(raw_data, recovered_data)
def selfplay(selfplay_dirs, selfplay_model, params):
"""Perform selfplay with a specific model.
Args:
selfplay_dirs: A dict to specify the directories used in selfplay.
selfplay_dirs = {
'output_dir': output_dir,
'holdout_dir': holdout_dir,
'clean_sgf': clean_sgf,
'full_sgf': full_sgf
}
selfplay_model: The actual Dualnet runner for selfplay.
params: A MiniGoParams instance of hyperparameters for the model.
"""
with utils.logged_timer('Playing game'):
player = selfplay_mcts.play(params.board_size, selfplay_model,
params.selfplay_readouts,
params.selfplay_resign_threshold,
params.simultaneous_leaves,
params.selfplay_verbose)
output_name = '{}-{}'.format(int(time.time()), socket.gethostname())
def _write_sgf_data(dir_sgf, use_comments):
with tf.gfile.GFile(
os.path.join(dir_sgf, '{}.sgf'.format(output_name)), 'w') as f:
f.write(player.to_sgf(use_comments=use_comments))
_write_sgf_data(selfplay_dirs['clean_sgf'], use_comments=False)
_write_sgf_data(selfplay_dirs['full_sgf'], use_comments=True)
game_data = player.extract_data()
tf_examples = preprocessing.make_dataset_from_selfplay(game_data, params)
# Hold out 5% of games for evaluation.
if random.random() < params.holdout_pct:
fname = os.path.join(selfplay_dirs['holdout_dir'],
output_name + _TF_RECORD_SUFFIX)
else:
fname = os.path.join(selfplay_dirs['output_dir'],
output_name + _TF_RECORD_SUFFIX)
preprocessing.write_tf_examples(fname, tf_examples)
def test_serialize_round_trip_no_parse(self):
np.random.seed(1)
raw_data = create_random_data(10)
tfexamples = list(map(preprocessing.make_tf_example, *zip(*raw_data)))
with tempfile.NamedTemporaryFile() as start_file, \
tempfile.NamedTemporaryFile() as rewritten_file:
preprocessing.write_tf_examples(start_file.name, tfexamples)
# We want to test that the rewritten, shuffled file contains correctly
# serialized tf.Examples.
batch_size = 4
batches = list(
preprocessing.shuffle_tf_examples(1000, batch_size,
[start_file.name]))
# 2 batches of 4, 1 incomplete batch of 2.
self.assertEqual(len(batches), 3)
# concatenate list of lists into one list
all_batches = list(itertools.chain.from_iterable(batches))
for _ in batches:
preprocessing.write_tf_examples(rewritten_file.name,
all_batches,
serialize=False)
original_data = self.extract_data(start_file.name)
recovered_data = self.extract_data(rewritten_file.name)
# stuff is shuffled, so sort before checking equality
def sort_key(nparray_tuple):
return nparray_tuple[2]
original_data = sorted(original_data, key=sort_key)
recovered_data = sorted(recovered_data, key=sort_key)
self.assertEqualData(original_data, recovered_data)